Merge branch 'attn-bwd-dropout' into fwd-drop-verify2

example/32_batched_gemm_scale_softmax_gemm/run_grouped_multihead_attention_forward.inc

@@ -10,7 +10,8 @@ int run(int argc, char* argv[])
     bool input_permute  = false;
     bool output_permute = true;
 
-    float p_drop                    = 0.1;
+
+    float p_drop                    = 0.2;
     float p_dropout                 = 1 - p_drop;
     uint16_t p_dropout_in_16bits    = uint16_t(std::floor(p_dropout * 65535.0));
     float rp_dropout                = 1.0 / p_dropout;
@@ -251,6 +252,7 @@ int run(int argc, char* argv[])
                           {seed, offset}); // dropout random seed and offset, offset should be
                                            // at least the number of elements on a thread
 
+
     // specify workspace for problem_desc
     DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
 

include/ck/tensor_operation/gpu/device/impl/device_grouped_multihead_attention_forward_xdl_cshuffle.hpp

@@ -623,6 +623,7 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle
                 //    z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5;
 
                 const auto z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+
                     GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
                         z_grid_desc_m_n);
 

include/ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_forward_xdl_cshuffle.hpp

@@ -1019,6 +1019,7 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle
                     z_thread_copy_vgpr_to_global.MoveDstSliceWindow(
                         z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
                         make_multi_index(0, 1, 0, 0, 0, 0, 0, 0, 0, 0));
+
                 }
                 else
                 {